Improve build system to avoid calling pkg-config many times (apache#4111)

* Improve build system to avoid calling pkg-config many times

* support specify cuda arch

Author: howard0su

Makefile

@@ -47,8 +47,8 @@ endif
 
 # setup opencv
 ifeq ($(USE_OPENCV), 1)
-	CFLAGS += -DMXNET_USE_OPENCV=1 `pkg-config --cflags opencv`
-	LDFLAGS += `pkg-config --libs opencv`
+	CFLAGS += -DMXNET_USE_OPENCV=1 $(shell pkg-config --cflags opencv)
+	LDFLAGS += $(shell pkg-config --libs opencv)
 	BIN += bin/im2rec
 else
 	CFLAGS+= -DMXNET_USE_OPENCV=0
@@ -183,15 +183,15 @@ build/src/%.o: src/%.cc
 
 build/src/%_gpu.o: src/%.cu
 	@mkdir -p $(@D)
-	$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" -M -MT build/src/$*_gpu.o $< >build/src/$*_gpu.d
-	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" $<
+	$(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -Xcompiler "$(CFLAGS)" -M -MT build/src/$*_gpu.o $< >build/src/$*_gpu.d
+	$(NVCC) -c -o $@ $(NVCCFLAGS) $(CUDA_ARCH) -Xcompiler "$(CFLAGS)" $<
 
 # A nvcc bug cause it to generate "generic/xxx.h" dependencies from torch headers.
 # Use CXX to generate dependency instead.
 build/plugin/%_gpu.o: plugin/%.cu
 	@mkdir -p $(@D)
 	$(CXX) -std=c++11 $(CFLAGS) -MM -MT build/plugin/$*_gpu.o $< >build/plugin/$*_gpu.d
-	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" $<
+	$(NVCC) -c -o $@ $(NVCCFLAGS) $(CUDA_ARCH) -Xcompiler "$(CFLAGS)" $<
 
 build/plugin/%.o: plugin/%.cc
 	@mkdir -p $(@D)
@@ -200,8 +200,8 @@ build/plugin/%.o: plugin/%.cc
 
 %_gpu.o: %.cu
 	@mkdir -p $(@D)
-	$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CFLAGS) -Isrc/operator" -M -MT $*_gpu.o $< >$*_gpu.d
-	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS) -Isrc/operator" $<
+	$(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -Xcompiler "$(CFLAGS) -Isrc/operator" -M -MT $*_gpu.o $< >$*_gpu.d
+	$(NVCC) -c -o $@ $(NVCCFLAGS) $(CUDA_ARCH) -Xcompiler "$(CFLAGS) -Isrc/operator" $<
 
 %.o: %.cc
 	@mkdir -p $(@D)

make/config.mk

@@ -48,6 +48,13 @@ USE_CUDA_PATH = NONE
 # whether use CuDNN R3 library
 USE_CUDNN = 0
 
+# CUDA architecture setting: going with all of them.
+# For CUDA < 6.0, comment the *_50 lines for compatibility.
+CUDA_ARCH := -gencode arch=compute_30,code=sm_30 \
+		-gencode arch=compute_35,code=sm_35 \
+		-gencode arch=compute_50,code=sm_50 \
+		-gencode arch=compute_50,code=compute_50
+
 # whether use cuda runtime compiling for writing kernels in native language (i.e. Python)
 USE_NVRTC = 0
 

make/osx.mk

@@ -45,6 +45,13 @@ USE_CUDA = 0
 # USE_CUDA_PATH = /usr/local/cuda
 USE_CUDA_PATH = NONE
 
+# CUDA architecture setting: going with all of them.
+# For CUDA < 6.0, comment the *_50 lines for compatibility.
+CUDA_ARCH := -gencode arch=compute_30,code=sm_30 \
+		-gencode arch=compute_35,code=sm_35 \
+		-gencode arch=compute_50,code=sm_50 \
+		-gencode arch=compute_50,code=compute_50
+
 # whether use CUDNN R3 library
 USE_CUDNN = 0